[Fix] Skip HF Hub API calls for cached repos and honor --revision

[mmangkad]

Motivation

Loading a cached model still calls the HF Hub API: transformers' revision lookups (new in 5.18) and repo listings, sglang's own checks, and dataset loads. In CI this exhausts the shared api rate limit (429). Several loads also ignored --revision and read main.

Merged main after #43229 landed; conflicts resolved as follows:

In main	Used	Why
Engine offline hook	Main, plus this PR's revision and draft handling	Covers in-process Engine startups, which this PR did not
Qwen3.5 shard-index fix	Main	Not covered here; correct as is
refs/<revision> lookup	This PR	Same logic, but main copies it inline for cache_dir; this PR reuses find_local_repo_dir
Weight-listing fallback	This PR (429 retry)	Main's fallback lists the local snapshot, which at that point usually holds no weights, so every allowed weight format is downloaded
Offline check, --revision	Main, with this PR's fix	Main ignored --revision, so revision-pinned models (e.g. Inkling test) never went offline
Offline check, draft models	Main, with this PR's fix	Main's Engine hook required tokenizer files for drafts, which they usually lack, so in-process speculative startups never went offline

Changes

Area	Change	Scope
Launch checks	GGUF, Mistral-format and DeepSeek FP8 checks read cached files before calling the Hub	All
Single files	LoRA configs, bitsandbytes, the token map and Inkling's quant config no longer list or snapshot the repo up front; only a Hub 404 or offline miss counts as a missing optional file	All
429s	Weight listing, snapshot downloads and diffusion Hub calls wait out a 429	All
Revisions	Inkling's quant config, the processor's tokenizer reload and the tokenizer helpers use the requested revision; get_tokenizer / get_processor take one revision	All
Eval data	LongBench-v2 and MMMU load from commit-pinned cached files	Tests
transformers lookups	Revision lookups and repo listings served from the local cache (get_cached_repo_tree)	CI
CI offline check	Subprocess and Engine paths honor --revision and check drafts by config and weights only; cross-repo auto_map code is accepted	CI

Verification

HF Hub api requests per launch (server + client), CI mode, warm cache. Outputs identical.

Launch	main	this PR
Qwen3-8B + KL helper (LongBench-v2)	38	0
Inkling --revision test, direct	67	0
Inkling --revision test, CI harness	67 (online)	0 (offline)
Llama-3.1-8B + LoRA	48	0

LongBench-v2 and all 30 MMMU subjects match the previous loads.

---

CI States

Latest PR Test (Base): ✅ Run #37894247607
Latest PR Test (Extra): ❌ Run #37894247194
Latest PR Test (AMD ROCm 10): ❌ Run #37894247523